Google DeepMind has introduced Genie 4, a real-time world model capable of simulating interactive environments frame by frame. Unlike traditional video generators, this system processes user inputs in real time to update the simulation. It is the first model from DeepMind to natively generate audio, music, and dialogue alongside visual content.
The architecture of Genie 4 integrates two components from the Gemini family. Gemini Omni defines the visual appearance and behavior of the environment, while other Gemini models manage logical coherence. Between user interactions, the system uses Gemini to steer the scene, ensuring continuous progression. DeepMind describes this as a hybrid approach where language models handle narrative structure while Genie renders output.
Genie 4 succeeds the Genie 3 model, which was announced in August 2025. While Genie 3 operated at 720p resolution with a frame rate of 24 frames per second and a visual memory of one minute, DeepMind has not provided specific benchmark tables for the new iteration. The company notes that it conducted internal safety evaluations prior to the preview release.
All audio and video generated by the model will include SynthID watermarks. Currently, the model is being tested by professionals across various sectors, including gaming, education, journalism, engineering, and creative industries.
